Hyperspectral Data Processing: Algorithm Design and Analysis

I

Preliminaries

PART I provides readers with preliminary knowledge and basic background to make this book self-contained. It comprises five chapters.

Chapter 1 covers fundamentals of subsample and mixed sample analyses. Chapter 2 addresses one of the challenges in hyperspectral imaging, that is how to deal with subpixels and mixed pixels often encountered in hyperspectral imagery. Since hyperspectral data are not necessarily image data, more generic terms, subsamples and mixed samples instead of subpixels and mixed pixels, are used to indicate that sample data can be either image pixels or spectral signatures.

Chapter 3 develops a new evaluation tool, a 3D ROC analysis, that extends the detection performance-based 2D receiver operating characteristics (ROC) curves commonly used in detection theory to measure estimation performance. As described in Chapter 2, most hyperspectral imaging techniques are developed as estimators rather than detectors to estimate abundance fractions of target substances for various tasks such as discrimination, detection, classification, etc. As a result, target detection that makes hard decisions is actually performed by target estimation that makes soft decisions in which case the 2D ROC analysis is not applicable unless these real-valued abundance fractions are thresholded to make hard decisions. The 3D ROC analysis is developed to meet this need by including a third dimension to convert a soft-decision-based estimator to a hard-decision-based detector. ...